J. Chem. Znt Comput. Sci. 1993,33, 763-768
763
Storage Techniques and Analysis Tools Founding a Knowledge-Based System for Conformation Prediction? Elke Lang’ Deutsches Krebsforschungszentrum, Zentrale Spektroskopie, Im Neuenheimer Feld 280, 69120 Heidelberg, FRG Jiirgen Brickmann Technische Hochschule Darmstadt, Institut fur Physikalische Chemie, Petersenstrasse 20, 64287 Darmstadt, FRG Received March 8, 1993 A conformation analysis system has been created that operates on the basis of a relational database network containing spectroscopic, crystallographic, and substance data, where a similarity-oriented structure retrieval method allows three-dimensional structures of the crystallographic database to be found by graphical input of a two-dimensional structure template and specification of the required degree of similarity. The set of hits is submitted as a whole to geometric conformation analysis by indicating the structure region which is common to all of the hit structures and specifying the positions and kinds of internal coordinates within the common region that should be computed. Results can be tabulated in files for further interpretation. This conformation analysis system helps in finding convenient starting structures for molecular modeling and in revealing the conformativevariability of specified structure classes for the detection of leads for the conformation prediction of similar structures in question.
INTRODUCTION Three-dimensional structure elucidation of biologically active compounds by means of spectroscopic methods and, related with this challenge, the detection of structure activities and computer-aided drug design using molecular modeling methods1 have become very important working fields. Rapid developmentin both concerned areas, NMR spectroscopyand molecular modeling, has opened new application possibilities. One more stage in this improvement is the use of structure database techniques for conformation retrieval and analysis, which is practiced at the Department of Spectroscopy at the German Cancer Research Center (DKFZ). Traditionally, molecular modeling methods have been applied to singular structuresfor the investigation of their conformationalbehavior and related properties. Soon it turned out that a collection of molecular structures with reliable 3D information would be able to support molecular modeling studies in many regards if it could provide flexiblesubstructure search and appropriate conformation analysis facilities. Very often, the Cambridge Crystallographic Data Files2 as the largest collection of experimentally obtained 3D structure coordinates are chosen as 3D structure source. For the intensive and fruitful use of an additional data source, however, it is indispensable to integrate it into the existing database environment. Only a homogeneous data access which spans all involved databases can provide synergistic effects and quality enhancement of the network of databases as a whole. Therefore, 3D structure access must be consistent with existing structure retrieval techniques in order to profit from their efficiency. Current methods for the detection of pharmacophoric patterns and the analyzation of conformational freedom in 3D databases3gain atom coordinatesby searching in a database of three-dimensional structure information with geometric algorithms. Database searching based on geometric properties needs the derivation of “bit screens- (description of the 7 This work is part of the dissertation by Elke Lang, TH Dannstadt, D17.
0095-2338/93/ 1633-O763$04.OO/O
structure features in form of a bit map) for the sake of fast search and conformative flexibility of the queries. These bit screens must be produced once by a time-consuming process before the database can be used for structure retrieval. The definition of 3D bit screens is very crucial and can be a limiting factor for completenessand exactness of the result. Therefore, a broad range of 3D bit screen sets in combination with various geometricsearch algorithmshave been reported. The database search is performed in at least two stages. The first stage is a fast screenout of entries that do not match the required bit screen pattern. The second and further steps are timeconsuming geometric analysis algorithms applied to the coordinate set of each of the structures which have passed the first step. Bit screen sets can be derived from atom distance distributions of calibration data! Other approaches tabulate the chemical function (e.g. H donor/acceptor, charge) of certain substructures as bit screensSor relate to the topological nature of the molecular structure.6 The first stage can be followed by a distance screenout’ or directly by geometric comparison of the required coordinate pattern with the coordinate sets of the entries. Several types of a1gorithms”ll are used to detect the target substructure in the possible hit structures. The remaining problem is that the two similar coordinatesets of query pattern and hit entry do not necessarily correspond to the same set of atoms linked together by chemical bonds.12 This problem must be overcome by estimating the degree of similarity between the two sets of coordinates.Several types of similarity estimation are compared by Pepperrell and Willett.” The aim of the present study was to overcome the various problems which have been described above to detect the desired target substructures in the possible hit structures which may become quite large. Our main strategy is the 3D structure access using 2D structure information. This way requires mapping of the 2D and the 3D structure information during the procedure of structure access and analysis of conformational freedom. This special approach offers the advantage Q 1993 American Chemical Society
764 J. Chem. Inf: Comput. Sci., Yo/. 33. No. 5. 1993
LANQAND BRICKMANN
of using the available 2D structure retrieval features directly. The amount of structures which have to be investigated by timeconsuming conformation analysis algorithms is dramatically lower than that which is produced by screenout with known methods. In contrast to screen-based systems, the presented method will not find any false positives. MATERIALS AND METHODS The system which will be outlined has been built on the basis of the database network of the DKFZ Department of Spectroscopy, which contains spectroscopic, crystallographic, andsubstancedata. This network provides substructuresearch facilities and, due to its structure-oriented design, offers the possibility of structure identification and structureoriented information linkage. The requirements which should be fulfilled by the system are raised by its purposes, structure elucidation in spectroscopy, and support of molecular modeling: (i) substructure search according to user-specified degrees of similarity, (ii) 3D structure access of the retrieved structure entries, and (iii) automatic conformation analysis performed on the set of structures without the need of making theatomnumbering ofsingleentriesconsistentto thereference structure by the user himself. Inthefollowing,themost importantaspectswillbedescribed in detail. This includes preconditions as well as special implementations. Database Techniques, Structure-OrientedInformation Network. In very clear contrast to most of the known databases in chemistry, the presented system has not been implemented as a dedicated file-management system but by means of a relationaPdatabase system product, SQL/DS, which provides features for the creation and maintenance of databases, input of data items, and property-oriented retrieval. This strategy assures that further databases can he added easily, using or enhancing the existing user surface and joining the contents of different databases by specifying their common properties (e.g. structure information for databases in chemistry). Database administrators and application programmers do not have to deal directly with questions of physical storage representation, since a data description language (DDL) is available for database conception and retrieval planning such as creating indexes for fast access. The data manipulation language (DML) ISQL is used to formulate queries by describing thedesired properties. Avoiding programming and maintenance overhead is a valuable help for scientists who, in fact, want to focus on the problems they want to solve with a database and not on the problems of database management. The information network of the DKFZ Department of Spectroscopy (Figure 1) has been founded with the spectroscopicdatabase SPEKTREN II,I5comprising a structure data table and several tables for experimental data. With the structuredatabase as the coreof the system, a toxicologic and a crystallographic'6 (Cambrige Crystallographic Data Files, CCDF) database have been added and a modeling database for the storage of structures which have been obtained by modeling techniques is currently developed. The structure database contains 2D structure descriptions in several degrees of similarity sharpness which serve to find structures which are identical or similar in constitution with a certain degree of similarity. Finding identical structures is very fast, as the structure-describing hash code derived using the Morgan algorithm'' serres as primary key for the data entries. Lower degrees of similarity slow the retrieval down within still acceptable ranges of time, as the display of results
Figure 1. Information network of the DKFZ Department of Spectroscopy containing spectroscopic, crystallographic, and toxicologic information. The different databases are joined together by acanonicaldescription ofthechemicalstructure. Key: KS, numerical key for the constitution as generated by the Morgan algorithm; KF, numerical key for the configuration as generated by the SEMA algorithm; SC, similarity code; F1,FZ, ..., stored information.
Figure 2. Three degrees of structuredescription accuracy: constitution, representation by the connation table (atom types. bonds and bond types); configuration, configuration of stereaanters indicated bydesciptorsderivedfrom 3D data;conformation,numeric valuesof torsion anglesconsideredas distinctive parameterskctween different conformations. already starts during the search process. A structure editor for the input of search structure templates and a specification menu which helps to indicate the desired degree of similarity prepare the information from which the SQL structure search argument is built. The results of structure search queries are presented at the screen (2D and 3D structure representation of each hit entry) with a colored indication of the search structure template in both structure representations. Dealing with 3D Information. Within the information network, a substance can be handled with three degrees of structuredescription accuracy (Figure 2) (structuredescription accuracy relates to different levels of description accuracy for one structure and should not be confused with the degrees of 2D structure similarity described above). The first degree is the description of the constitution (2D structure) according to the Morgan algorithm. Many of the database entries can possess the same constitution. One constitution description, if it shows at least one stereocenter, can lead to several configurations which can all be present in the database. Therefore, the configuration description is the second degree of accuracy. The third and highest degree of accuracy is the structure conformation, represented by the (3D) coordinates
KNOWLEDGE-BASED SYSTEMFOR CONFORMATION PREDICTION
I
alom- and bondwise matching
A
(search template. reference structure)
L
eombioation of
mappines
(region 10 be investigated during conformation analysis)
Figure 3. Conccpt of structure handling: 3D substructures in the
database are found by linkage of two mappings, atom- and bondwise matching and 2LL3D matching. of the structure. Conformation analysis can only lead to meaningful results if the geometric comparison is performed on structures that are similar at the lower degrees of accuracy (and hence comparable) within the structure region being under investigation. For this reason, observing the restriction that only structures with known topology can be handled, a special concept of structure handling has been developed (Figure 3): Any type of structure access, 2D, stereo-enhanced 2D, and 3D, begins at the 2D level as the access level and satisfies the desired degree of accuracy by mounting in accuracy as far as needed. This proceeding avoids geometric 3D search and the need of deriving 3D bit screens but offers access to 3D substructures via their corresponding 2D substructures. The correspondence between 2D and 3D structure representation (recognizing a certain atom of the 3D representation as corresponding to a certain atom of the 2D representation and vice versa) is found by a 2D to 3D matching procedure.18 The intermediate level of stereosensitive 2D description has not been made available so far for CCDF users with the original QUEST software. Within the described informationnetwork,stereochemistryis represented on the baseofthestereochemicallyextendedMorgan algorithm (SEMA)19foralloftheinvolveddatabases.FromtheSEMA representation and the 3D coordinates, an R/S indication of stereocenters according to Cahn-Ingold-Prelog (CIP) nomenclature” can be derivedz1and offers a description which is familiar to most of the users. The stereodescription of structures offers the possibility of stereoselective structure retrieval and comparison. Conformation Analysis. Structure data which have been extractedfromthedatahasecanbesubmitted toconformation analysis. A set of structures bound for investigation must possess a structure region in common. This common substructure is specified in the graphics-oriented user menu and forms the hasis for analysis requests. Distances, bond angles, and torsion angles are computed for the atom sets and operations which have been required by the user by means of the menu. If stereocenters are detected in the common structure region, the set of structures is partitioned into the resulting classes of different stereoisomers which contribute to different sets of results. The analysis results are listed by
J. Chem. In$ Comput. Sci.. Vol. 33, No. 5. 1993 765
using the numbering scheme of the common substructure and indicating its relationship to the numbering scheme of the individual entries so that the derivation of histograms and distributions, but also a closer look to certain single entries, is possible. Details about the proceeding of the conformation analysis module have been reported in ref 22. The conformation analysis feature serves as a tool for structure comparison and evaluation of the effects of ZD structure similarity on conformative similarity. The second challengeofconformationanalysis is the derivation of numeric data (internal coordinates, mainly torsion angles) for a conformationdescription. Thecurrent implementationof the system provides the possibility of operating on the conformations of structures which have been retrieved. The further step which is now being implemented is a conformation description which can be searched directly. Here again the benefits of using a relational database system product must be stressed the flexibility of the database architecture and the system features for creating and modifying data tables or data rows in existing tables makes it easy to add data fields to the existing data tables and to create test versions of conformation description concepts. RESULTS APPLICATION OF THE SYSTEM TO THE INVESTIGATION OF PHOTOSENSITIZING PORPHYRIN DERIVATIVES Similarity-based substructure retrieval and conformation analysis techniques as described above have been applied to several problems in the field of computer-aided drug design which arose during the work with medicinal chemists and pharmacologists. One of these cases will be outlined below in order to demonstrate how the questions posed by users of computer-aided drug design techniques can be satisfied by the system. Porphyrins are used as photosensitizers for the photcdynamic cancer therapy and in tumor diagnosis, as porphyrinenriched tumor tissue produces characteristic fluorescence effects. Several properties of porphyrins must be optimized to make them satisfying tools in cancer therapy: Photosensibilization must concern a light spectrum which does not overlap with normal daylight spectra as patients should be able to be exposed to daylight during therapy. Enrichment in tumor tissue should be high, whereas enrichment in normal tissue, especially in liver and kidneys, should be minimal to avoid necrosis of tumor-adjacent tissue during irradiation. Tetraphenylporphyrinsnormally do not satisfy these conditions as their water solubility and molecular weight are too low. A possible solution is to increase the molecular weight and the volume of the structure by adding further substituents to the phenyl groups. The abovementioned therapeutic requirements can be translated into terms of structure chemistry: (1) Are the phenyl rings conjugated with the porphyrin skeleton? (2) Are the four phenyl substituents rotatable or ‘frozen” in their positions? (3) Which constellations of torsion angles do the phenyl rings exprime, and which frequenciesof constellation patterns can be observed? (4) Whichconformationsof theporphyrinskeletondoexist? How does the skeleton adapt to central ions of different sizes? The CCDF were searched for tetraphenylporphyrins by substructure retrieval (Figure 4). A set of hits with complete and reliable structure information (52 structures which were
766 J. Chem. In5 Comput. Sei., Yo/.33, No.5. 1993
LANGAND BRICKMANN
-
generation of structure similarity codes
SEMA structure
BUTVOB
% classification of conformers
d
P
D
interpretation of results, derivation conformation description
4. Information flow in the system. After input of the desired structure. structure description d e s are derived and used as search argumentsforthcsvucturedata~sc.Foreachofthchit cntries,slnrcturc-rclatedinformationisfoundintheCCDFdataublesviathcSEMA code linkage hey. Conformation analysis and interpretation of the results lead to the derivation of conformation descriptions, which can k stored in the modeling databasc.
suicablcaocordingtosubstitutiontypcand size) wassubmitted toconformationanalysis. Theconformationparameterswhich were chasen to be computed were tabulated and allowed to estimate plausible ranges for the concerned parameters. Considerationof characteristic differences of torsion and bond angles allowed, by numeric criteria, detection of the four distinct shapes of the porphyrin skeletons which had already been reported by Hamor et aL23 The distinctive parameters and their values are given in Figure 5 and Table I. The above questions can be answered as follows: (1) Torsion angles between 80' and 120" and bond length vaiu& between 129 and 1.51 A show that phenyl groups and porphine skeleton are not conjugated.
(2) Visual consideration of the van der Waals radii gives an evident pmof that free rotation of the phenyl groups is impossible for geometric reasons. (3) The orientations of the phenyl groups toward the porphineskeleton(defincdaspositiveornegativetorsion s e w of the four torsion angles from the porphine skeleton to the phenyl groups) is nearly equally distributed on the sequence patterns 'four equal signs", 'alternating signs", alternating", and 'three equal signs" and implies that the orientation of the ring closure constellation is conserved in the structure shapc. (4) The conformation types could be detected applying the numeric criteria listed in Table I. The question of the adaption
++
J. Chem. Inf. Comput. Sci., Vol. 33, No. 5, 1993 767
KNOWLEDGE-BASED SYSTEM FOR CONFORMATION PREDICTION
Table I. Distinctive Conformation Parameters Which Were Used for the Detection of Porphine Skeleton Shapes (See Figure 5 for Localization of the Angles)' i ii iii iv I1 I11 N-Met-N i ii iii iv I1 111 N-Met-N A B
7.4 7.4
170.6 170.6
5.3 5.3
174.3 174.3
175 175 180
175 175 150
180
